Researchers have developed SlimDT, a modification of the Decision Transformer (DT) model for offline reinforcement learning. SlimDT removes the Return-to-Go (RTG) token from the autoregressive sequence, instead injecting this information directly into the state representations. This approach reduces the sequence length by one-third, leading to improved inference efficiency and computational gains. Experiments on the D4RL benchmark show SlimDT outperforming standard DT and achieving performance comparable to state-of-the-art methods. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Introduces a more efficient variant of Decision Transformer, potentially improving performance and reducing computational costs in offline reinforcement learning tasks.
RANK_REASON This is a research paper detailing a novel modification to an existing model architecture for reinforcement learning. [lever_c_demoted from research: ic=1 ai=1.0]